Breaking Bias: A Context-Aware Multimodal Framework for Detecting and Neutralizing Ideological Bias in News

Media bias, particularly bias by word choice , influences public perception and fosters polarized clusters of opinion. Biased word choices, often subtle and context-independent, erode trust and often jeopardize social frameworks. Detecting and neutralizing such biases remains a challenge, as existing feature-based and neural models fail to capture nuanced ideological leanings, especially in dynamic news contexts. This issue has been addressed by proposing a multimodal framework CANDOR ( Context-Aware Analysis for Neutrality and Detection of Objective Reporting ) for detecting and mitigating media bias. CANDOR integrates ideology-aware text representations and context vectors from visual and textual data in order to improve bias detection and neutralization effectiveness. The framework leverages BIASLENS , a novel pre-training objective that contrasts same-story articles of varying ideologies, and VT-CT , a vision-text context transformer to enrich contextual embeddings. It detects polar sentences, extracts ideological biases, and generates neutralized content while preserving factual accuracy. CANDOR employs semantic loss functions targeting fluency, content preservation, and bias neutralization for natural language generation. Tested on benchmark datasets like ALLSIDES , the framework demonstrates a \(4\text{--}13\%\) improvement in accuracy and \(1\text{--}12\%\) enhancement in content preservation , outperforming state-of-the-art models. Human evaluations confirm the framework’s ability to neutralize media bias while maintaining readability. This work provides a step toward mitigating media bias and fostering balanced, objective news reporting.